Strategy

Who pays for the last mile in enterprise AI?

Editorial · Reveneau · September 19, 2026

Who pays for the last mile in enterprise AI?

Wonderful's numbers do something no enterprise AI company has done at this size before: they put a price on the last mile in public. On 6 September 2026 CTech's Sophie Shulman reported that the company keeps a gross margin of about 52 percent of revenue, against the 70 to 90 percent a software company usually keeps, and that about 400 of its 650 employees work at customer sites. Globes had reported a month earlier that headcount went from 90 to around 630 in one year.

Gross margin is the share of each dollar of revenue left after paying the direct cost of delivering the work. The 18 to 38 cents Wonderful gives up on each dollar, compared with a software company, is mostly the salaries of the people it sends into banks and telcos to get AI running. That is the last mile, priced.

We have built for clients long enough to know that this bill always exists. A model in a demo and a system in production are separated by integration, permissions, security review, and the work of getting people to change what they do on Tuesday. Someone pays for that every time. What's changed is that we can now see who, at scale, and there are only three ways the bill is ever settled.

1. The vendor pays, out of margin and out of its investors' money

This is the way Wonderful is paying today. Its founder Bar Winkler described the work to Globes in June 2026: "We need to connect systems, understand processes and persuade organizations to change the way they work". The people who do that are on the company's payroll, and the margin shows it.

The money to run this way comes from investors, and the investors have made the argument for it in writing. Andreessen Horowitz's Joe Schmidt published an essay on 4 June 2025 titled Trading Margin for Moat. ServiceNow, he noted, had a gross margin of 63.2 percent at its IPO and later reached 79 percent. Workday started at 54.1 percent and reached 75 percent. The implementation work that cost the margin early is what tied customers to the product, and a founder who optimises for an 80 percent margin from day one, in his view, gives that up. Wonderful has raised more than $800 million, per CTech, which buys a long time to run at 52 percent.

Shulman's report adds the part a buyer should hear. The company can grow this fast, she wrote, partly because it prices its services low compared with Salesforce and Accenture, and she drew a parallel to Uber's early years, when venture capital subsidised prices before the company had to face the economics of operating at scale. For a buyer, a subsidised deployment is a good deal while it lasts. The question is what the renewal looks like once the vendor's investors expect software margins. Wonderful's chief technology officer Roey Lalazar has answered that in advance: "leaving Wonderful should be easy", every agent and configuration can be exported, and, in his words, "We cannot raise prices arbitrarily." That is the right answer. Get it in the contract.

2. The customer pays, with its own people

The second way is the handover. Wonderful's homepage describes local teams who "build it with you, then hand it over and move to the next use case," and its Bank Hapoalim story describes a bank data engineer with no agent experience shipping the second agent in three weeks by reusing 15 tools the vendor's team had built, by the company's own account.

When that works, the bill for every deployment after the first moves from the vendor's invoice to the customer's payroll, which is where a customer usually wants it. The customer's people know the systems, the risk is theirs to manage, and the vendor's margin recovers because it is no longer paying for the second mile.

The trade is only fair if the capability actually moved. A handover that leaves a system nobody in the building can change has moved the bill without moving the ability to pay it, and the customer ends up calling the vendor for every change at whatever the vendor then charges. The test is a count: how many of your engineers can ship a change to the system, verified, without the vendor on the call? Run it on something small before the vendor leaves. Pick one real change, a new field in the intake form or a changed refund limit, and have one of your own engineers take it from ticket to production with the vendor watching and saying nothing. If it ships, and the checks that prove it ship with it, the capability moved. If the engineer needs a call in the first hour, the handover was a document, and we wrote about what a real one contains in What a good handover document actually contains.

3. The work itself gets cheaper, which is the only way that scales

The third way is the one the first two are betting on, and the one nobody has proved yet at Wonderful's size. If the second deployment reuses the integrations, the checks and the tooling the first one built, the marginal work falls each time, and the margin rises without anyone raising prices or handing the bill to the customer.

Reuse has a specific shape in the work. The first deployment into a telco writes the connector to its ticketing system and the checks that prove the connector behaves: a ticket is opened with the right fields, closed with the right status, never duplicated. The second deployment into the same kind of system starts with that connector and those checks already written, so its clock begins at the part that is new. Each deployment leaves the next one shorter, and the number that shows it is the count of checks the second deployment inherited before writing a single one of its own.

The venture capital analyst Eze Vidra put the test in one line on 3 September 2026: "fieldwork only becomes a moat when it becomes product. If every engagement starts from zero, the startup is building a consultancy." A separate analysis from 3V made the same point as a number, arguing that the figure to watch in Wonderful's results is revenue divided by headcount, because it shows whether each new customer needs the same number of people as the last. Wonderful's own bet on this is its Agent Builder, a product that builds and tests other agents, which it says cut build times by up to 50 percent across more than 60 deployments, by its own account.

This third way is our whole method, so we should say plainly how we pay for the last mile. AI writes the code. An eval suite written from the specification proves every change before it ships. That lets a small team do the deployment work a large one would otherwise do, and the saving goes into the client's price instead of into our margin. We describe that as the mechanism we work by. No saving has been measured yet, and a number without the working behind it is the kind of claim this whole piece is about. Our guide to the model, and to when it is the wrong answer, is at forward deployed engineering.

So the question to ask any vendor whose engineers will sit in your building, including us, is the same one Vidra asks Wonderful. Did your second deployment for a comparable customer take fewer people and fewer weeks than your first, and can you show the working? A vendor whose answer is a bigger team is selling you the first way. A vendor whose answer is a reused eval suite and a shorter clock is selling you the third.

Our thanks to Sophie Shulman at CTech and Eze Vidra at VC Cafe, whose reporting made a cost that is usually hidden inside a services line visible enough to write about. The last mile never gets free. It only gets cheaper, or it gets hidden.

Sources

Common questions

What is the last mile in enterprise AI?

It is everything between a model that works in a demo and a system serving real customers inside a company's own stack: connecting to billing, identity and ticketing systems, deciding what the system may not do, passing security review, and changing how the people around it work. It is engineering and change work, and it costs money whoever does it.

What does a 52 percent gross margin mean in plain terms?

Gross margin is the share of each dollar of revenue left after paying the direct cost of delivering the work. CTech reported in September 2026 that Wonderful keeps about 52 cents of each dollar against the 70 to 90 cents a software company usually keeps, and the difference is mostly the salaries of the 400 people it has working at customer sites.

Why would a vendor accept a lower margin on purpose?

To make its software hard to remove. Andreessen Horowitz's Joe Schmidt argued in June 2025 that ServiceNow and Workday both went public with gross margins in the 50s and 60s and reached the 70s later, because the implementation work that cost margin early is what tied customers to the product.

Is a subsidised deployment a good deal for the buyer?

Yes, while it lasts, and the question is what happens to the price when it stops. Ask what the renewal looks like once the vendor's investors expect software margins, and whether you can leave; Wonderful's chief technology officer has stated that leaving should be easy and that the company cannot raise prices arbitrarily, which is the answer you want in writing.

What does it mean when the customer pays for the last mile?

It means the vendor hands over the system and the customer's own engineers ship the next workflow, so the ongoing cost moves to the customer's payroll. That is a fair trade only if the capability actually moved, and the test is how many of your people can ship a change without calling the vendor.

How can the last mile itself get cheaper?

By making the second deployment reuse what the first one built: the integrations, the checks, and the tooling, so the marginal work falls each time. The venture capital analyst Eze Vidra put the test in one line in September 2026, that fieldwork only becomes a moat when it becomes product, and if every engagement starts from zero the company is a consultancy.

What single number tells you which way a vendor is going?

Revenue divided by headcount, and whether it rises as the customer count grows. A 3V analysis of Wonderful's Series C argued that this ratio is the figure to watch, because it shows whether each new deployment needs the same number of people as the last.

How does Reveneau pay for the last mile?

By having AI write the code and an eval suite written from the specification prove it, so a small team does the deployment work and the saving goes into the client's price instead of the margin. We describe this as the mechanism we work by, because no saving has been measured yet and we do not print numbers we cannot show the working for.

What should I ask an AI vendor about the cost of its second deployment?

Ask whether its second deployment for a comparable customer took fewer people and fewer weeks than the first, and ask to see the working. A vendor whose answer is a bigger team is selling you the first way of paying; a vendor whose answer is a reused eval suite and a shorter clock is selling you the third.